4 The determinants of migration and remittances in
4.3.2 Methodology
The methodology for the determinants of migration is straightforward. The determinants of being a migrant (a dichotomous variable taking on values of 1 for being a migrant and 0 for not being a migrant) are measured. I follow the previous literature by using a standard probit model, since I want to analyse the selection process. Equation 4.1 outlines the probability of being a migrant. 46 ) ( ) 1 (y= x =G β1x1+β2x2+β3x3+β4x4+β5x5+β6x6+ν p 4.1 G is a function taking on values between zero and one, and since this is a probit model it is a standard normal cumulative density function. p measures the
probability of an individual to be a migrant, given the following characteristics. x1 are individual characteristics. The main variables of interest are regressors x2 to x6. These are characteristics that refer to the different parts of the Welfare Pentagon. x2 are family characteristics, x3 are market characteristics, x4 are membership institution characteristics, x5 are social network characteristics and x6 are public
l authority characteristics. ν is the error term referring to any unexp ained differences.
Finding proxies that refer to different parts of the Welfare Pentagon is not a straightforward task. For example, how does one model whether someone has a helpful family? In this instance, I use family size (extent of possibility of help) and family wealth (extent of their helpfulness) as a proxy. In the next section I outline in detail which proxies are used for which corner of the Welfare Pentagon. Due to data limitations (especially for Moldova), the proxies used are quite rough. Nevertheless these proxies should give a basic idea about the different parts of the Welfare Pentagon.
The methodology for the determinants of remittances requires more explanation. Early papers on the motivations to remit used Ordinary Least Squares (OLS) (for example Lucas & Stark, 1985) to model the remittance decision. It is now known that using such a method leads to biased and inconsistent estimates, since a substantial fraction of the migrants does not remit. In recent papers, the main methodological distinction is made between modelling the motivations to remit as a one‐stage decision (Tobit) where the decision to remit and the amount of remittances are made together or as a Heckman two‐stage approach (Probit and corrected OLS) where the model separates the decision to remit and the subsequent decision of how much to remit. The advantage of the latter approach is that it allows a regressor to differently affect the decision to remit and the level of remittances. Amuedo‐Dorantes & Pozo (2006), on the other hand, argue that using
s
a two‐part selection model leads to identification problems, i.e. it is hard to ay which variables would matter for one decision and not the other.
An alternative to the two‐stage approach is to assume that there is only one remittances decision in which the two stages occur simultaneously. This one‐stage decision can be modelled as a single equation estimated by Tobit analysis, using both remitting and non‐remitting migrants. Each regressor has the same effect on the probability of being a remitter and on the level of remittances. The convenience of this approach is that it enables the identification of a set of variables that are most significant in influencing “remittance behaviour” and thus test some of the theoretical predictions which do not differentiate between the refore, assume that the remittance using a Tobit model. decision to remit and how much to remit. I, the decision is a one‐stage process and will model it The Tobit model is sp ⎪ 0 ecified as in equation 4.2 : 4.2 ⎪⎩0 where ⎨ ⎧ ≤ > = 0 * * * i i i i R if R if R R i i i X R*=β' +ε
The Tobit model is used for censored data, where the dependent variable R*i is latent. In the following analysis R*i is observed for values that are higher than zero and it captures the i‐th individual’s propensity to remit. It has a normal, homoskedastic distribution with a linear conditional mean. Ri is the actual observed value of remittances remitted by individual i. It can be either positive or zero and it is positive for those migrants that do remit. Xi is a vector of explanatory
48
variables that explains whether and how much someone remits. ε refers to the error term that accounts for any unexplained differences.
A disadvantage associated with the Tobit model is that the assumption of normally and homoskedastic distributed errors might not hold. If households have more than one remitter, remittances of both remitters partially depend on the same unobservable household characteristics and this results in error terms that are correlated across observations.30 Since most households in the datasets only have one remitter, I assume that this problem is minimal.31
4.4
Why do people migrate?
This section looks at why people migrate. It uses data from standard household surveys for Albania and Moldova to explore whether something can be said about the relevance of the Welfare Pentagon institutions in explaining differences between migrants and non‐migrants. It is clear that these data are not collected to test the livelihood portfolio theory.32 Many variables that would reflect the access to, endowment and investments in the institutions of the Welfare Pentagon are simply not available. This means that I roughly explore whether (not) having access to certain parts of the Welfare Pentagon explains why some migrate while others do not. Section 4.4.1 gives a first indication with descriptive statistics, followed by the econometric analysis in section 4.4.2.